Mitigating Input Prompt Attack Vulnerabilities in Systems with a Language Model Interface

The article is devoted to the study of the cybersecurity vulnerabilities in systems utilizing Language Model Interfaces, particularly focusing on the challenges and solutions in constructing secure systems. The paper includes an overview of the current Language Model Interface and issues with their utilization in typical systems, and because of the conducted analysis, proposes a multi-level structural schema of cybersecurity frameworks and standards, that could be applied to cover the different types of risks. The suggested prompt analysis and injection detection sub-system design, in which evaluates whether the input is relevant and appropriate from the security point of view. The additional Filter cybersecurity level integrated into the system with the Language Model Interface allows protection from prompt injection attacks during the pre-processing of user requests and post-processing of responses. The Filter level consists mainly of a Prompt Analyzer, which wraps pre-processed input prompt; an Attack Validator module using a helper Large Language Model, which classifies if the prompt is risky or safe; Elasticsearch cluster for logging and monitoring purposes. During testing different types of prompt injection attacks were evaluated in different types of system setups: core GPT-3 and GPT-4 models without additional security layer; Fuzzy Search with attacking prompt pattern vocabulary; a system with Attack Validator in two modes - based on GPT-3 model and GPT-4 model. The article shows the relative strengths and weaknesses of each approach, contributing valuable insights into the development of more secure systems with Language Model Interface.